Back to tags
#prometheus
metrics, scrape configs, and the alert that fired at 3am for the right reason.
3 blog posts.
Posts on Prometheus as the thing that decides whether anyone gets woken up. The Grafana guide is the stack end to end, dashboards and alerts included. The burn-rate post replaces threshold alerts with recording rules over the error budget, plus the absent_over_time alert that covers the case where the exporter itself disappears and every other alert goes quiet. The GPU observability part is the same reasoning applied to model serving, where utilisation is the misleading metric and queue time is the honest one.
Blog posts
Threshold alerts page you for causes that may not matter. Multi-window burn-rate alerts page you once, for the symptom, when the error budget is genuinely…
The GPU dashboard says 92% busy and users are waiting eight seconds for the first token. Monitoring an LLM server means watching the queue, not the…
A practical guide to setting up Prometheus and Grafana for production monitoring. No theory, just battle-tested configurations that work.